Agents take the intelligence. People keep the judgment.
Every autonomy program eventually collides with the same question: which work goes to the agents, and which stays human? Most organizations answer it politically — by department, by seniority, by who objects loudest. There is a cleaner line, and it is the one that determines whether autonomy compounds or stalls.
Work divides into intelligence and judgment. Intelligence is work with a verifiable right answer: match this payment, code this invoice, test this control, translate this balance. It is rule-dense and high-volume, and its defining property is that correctness can be checked without asking anyone. Judgment is work where the answer depends on appetite: accept this variance, extend this credit, sign this filing. Its defining property is accountability — someone must own the consequence.
The division of labor is principled, not political: if correctness is checkable, it belongs to an agent under controls. If the consequence needs an owner, it belongs to a person.
Why the split matters economically
Assistance tools blur this line — a person still touches every unit of work, so cost scales with volume no matter how good the tool gets. Drawing the line properly breaks that scaling: intelligence runs at machine cost, and human cost concentrates on the exceptions where judgment is actually exercised. That is the difference between a productivity tool and a change in the cost structure of a function — and between per-seat pricing and paying for outcomes.
The exception queue is the new job
When intelligence moves to agents, the human role does not shrink — it concentrates. The controller who touched a thousand reconciliations now governs the twelve that fall out, each arriving with context, the agent’s proposed resolution, and the evidence trail. Three disciplines make this work. Exceptions must arrive with proposals, not just alerts — triage without a recommendation is the old job with worse tooling. Every decision must be captured as policy, not folklore, so the same exception never needs the same human twice. And the escalation path must terminate in named owners, because autonomy without an accountable human at the top is not an operating model; it is a liability.
The boundary moves — deliberately
Today’s judgment becomes tomorrow’s intelligence. Decisions repeated often enough, consistently enough, reveal their rule — and once the rule is explicit, the work crosses the line. This is what an autonomy dial governs: not a one-time delegation, but a ratchet, moved per process and per threshold, on evidence, with maker-checker on every promotion. Organizations that manage the boundary deliberately get compounding autonomy. Organizations that don’t get one of two failures: agents trusted with judgment they haven’t earned, or people trapped doing intelligence a machine mastered a year ago.
The end state is not fewer people doing the same jobs faster. It is a smaller number of more senior jobs: owners of thresholds, controls, and exceptions — the judgment layer of a function that otherwise runs itself.